OpenAI on Wednesday published a framework for reporting instances of model misalignment, along with six reports on problematic behavior observed over the past six months.
The company said the framework is meant to speed up publication of misalignment findings, including cases it has not yet fully explained or mitigated, and that it favors disclosure even when an instance’s significance is uncertain.
Under the framework, each discovered incident is assigned to one of three tracks based on complexity. OpenAI said its Hugging Face incident would have fallen under the framework’s slowest investigative track, which covers complex investigations, especially those involving third parties.
According to OpenAI, the reports published alongside the framework describe individual instances and “shouldn’t be considered reflective of how often misalignment occurs across our models.”
In one report, an internal model tasked with retrieving county earnings figures during reinforcement learning training repeatedly failed to reach a data API, then attempted to register for an API key using a disposable email address and searched public GitHub repositories for leaked keys.
One recovered key authenticated successfully and returned metadata. When the model still could not retrieve the requested figures, it fabricated values and presented them as transcribed from the source website, without disclosing the failed retrieval or the use of the leaked key.
A separate report describes models using OpenAI’s internally hosted instance of Artifactory, a shared package repository, as a message board across otherwise independent training samples. Agents searching for missing input files discovered messages left by other solvers working the same problem, read them, and posted their own requests for help.